[Request] Multithreaded imwrite

[weux]

When processing a large number of images I notice that my GPU will max out to 100% while it's processing, then drop to 0% while the output file is being written to disk. Is there any way to get the next image to start processing while the last image is being writen? This would make a big difference when processing an image sequence thousands of frames long.

[iame6162013]

May I ask what the slow part is?
Is it saving to disk that is slow, i.e. the disk is >80% utilized, or does the compressing e.g. to png take too long?

[iame6162013]

Is loading files slow also?

[weux]

I guess slow might not be the best way to describe it, but there is small <1 second delay between files. I think it's a combination of compressing the output png, saving it to disk, and uncompressing the next input png.

For just a few images this is no big deal, but if you try to process 100,000 images the small delay adds up.

Eliminating that delay between files would keep the GPU running at 100% the whole time. Then your processing time per image would be limited solely to your GPU, rather than GPU + storage performance.

[byron39]

I would like to confirm this is a image loading performance issue, and also provide more details I have found.

Using a GTX 1080 in OpenCL mode for all tests, an input resolution of 512x288, scale ratio of 2.5, and an output resolution of 1280x720, over a group of hundreds of images.
Setup 1:
With input and output storage of a 7200RPM HDD, waifu2x performs at 0.72 fps.
Setup 2:
With input of a RAM drive, output on the same 7200RPM HDD, waifu2x performs at 1.06 fps.

Suggestions based on these observations:

• After the first image is sent to the "processor" for up-scaling, have the program copy/load the next image (into RAM) so it is available before it is needed, effectively removing the disk latency.
• It seems any pre-processing for GPU/"processor" consumption could also be done ahead of time as well.
• Maybe also save the images after the GPU/"processor" has been fed work again, reducing job latency further.

Hope these observations help.

[nioroso-x3]

Using gnu parallel is another solution, for example with a folder full of pngs:
Create a list:
ls $(pwd)/*.png > list.txt
Use the list as input for GNU parallel
cat list.txt | parallel --progress --joblog job.log -j 4 "./waifu2x-converter-cpp -i {1} -o {1}.up.png"
With 8 threads I have full utilization of a fury nitro, 2400 gflops effective.